| name | sc:codebase-index |
| description | Semantic codebase search & indexing (Cursor-style). When a question is conceptual ('where is X handled', 'find everything related to auth') or the codebase is large/unfamiliar, index it once with the code-memory tools and use semantic search instead of blind grep — then verify hits by reading the real files. Activate for: where is this implemented, find where, search the codebase, semantic search, index the codebase, find all usages of, how does X work, איפה זה מטופל, חפש בקוד, אינדקס לקוד, חיפוש סמנטי, מצא את כל המקומות. |
| license | MIT |
| metadata | {"author":"squadcoder","version":"1.0.0"} |
Codebase indexing & semantic search (reuse the code-memory MCP)
SquadCoder ships the code-memory MCP (local, offline, no API key — uvx code-memory) for
Cursor-style semantic code search. This skill tells you when and how to use it so the agent indexes
and searches the codebase the way a Cursor user expects — without anyone wiring it by hand.
If the code-memory tools (index_codebase, search_code, search_docs, search_history) are
not available, the MCP is disabled — fall back to grep/glob and (optionally) tell the user they
can enable code-memory in Settings ▸ MCP for semantic search. Do not invent these tools.
When to use semantic search vs grep
- Conceptual / fuzzy ("where is rate-limiting handled?", "find everything about billing", "how does
auth flow work?") → semantic (
search_code / search_docs). It finds relevant code even when the
wording differs from the symbols.
- Exact token (a known symbol, string literal, error code, file name) →
grep/glob. Don't pay
for embeddings when you already know the literal.
- History / "who changed this and why" →
search_history (commit/blame/file-history).
Flow
- Index once per workspace. On the first semantic query in a session (or when the user opens a new /
large codebase), call
index_codebase for the project root. It's incremental — re-running is cheap and
keeps the index fresh after edits. The first ever run downloads the embedding model (~600 MB) and may
take a minute; say so if the user is waiting.
- Search. Use
search_code (and search_docs for prose/markdown) with a natural-language query.
Pull the top hits.
- ALWAYS verify before acting. Semantic hits are leads, not ground truth — open the actual files
and read the real lines before you edit, quote, or conclude. Never edit a file you only saw in a search
snippet.
- Keep it fresh. After a batch of edits that change structure, re-index so later searches stay
accurate.
Notes
- Fully local & offline — code never leaves the machine (fits SquadCoder's privacy posture).
- Prefer this over re-reading the whole tree when the codebase is big; prefer direct reads when it's small.
- This is a thin reuse layer over the MCP — no engine changes; works the same in CLI, desktop, and web.
